Source: raw/Nate_Herk_Breaks_Down_His_3_Most_Powerful_AI_Automations.md

Creator: Matt Wolfe (host) with Nate Herk, “Show Me Your Workflow” series, episode 2 · URL: https://www.youtube.com/watch?v=AtjFk9f44HE · Duration: not recorded · Platform: YouTube (fetched 2026-11-01)

Nate Herk walks Matt Wolfe through three Claude Code builds he uses every day: a video-editing skill trained on his own best-performing edits, a “Roast Room” council of famous-founder subagents that debate a decision, and a news Instagram account he says is 98-99% automated. All three are skills in a public GitHub repo; Wolfe’s install advice is to paste the repo URL into Claude Code and ask it to set it up. The captions render his name “Nate Hurt”; it is corrected here.

Key Takeaways

  • Build skills from proven work, not from the model’s taste. Herk had Claude Code analyse 27 of his best-performing human-edited reels frame by frame (scene changes, B-roll timing, shimmer effects, voice-versus-music levels), turned the result into “a bank of patterns”, and built the editing skill on that. He compares it to evals: start from the golden data set you are building toward.
  • The editing stack. Claude Code drives HyperFrames for the render, a browser for B-roll (screen recordings, clicks, screenshots), AI video generators for inserts, and Blender for 3D scenes he has never opened manually. Reels use three scenes: him at the bottom with B-roll on top, full-screen B-roll, and full-screen him.
  • Word-level timing from ElevenLabs. He transcribes with the ElevenLabs API because it is fast and cheap (local Whisper is free but slower). Word timestamps (e.g. a word spoken from 0.09 to 0.17 s) let motion graphics land on the word. To stop words being clipped, the skill also checks the audio waveform and cuts only where it goes flat.
  • Long-form editing rules he baked in. Switch to full-screen face cam when the screen recording has been idle (mouse still, likely a rant); add zoom-ins so viewers can read text; flag spots where his explanation leaves a gap; and keep an LLM-as-judge that improves as he gives feedback. It also follows pointing gestures to place logos.
  • Screenshot loops for anything visual. Claude Code cannot watch video, so for short-form it screenshots roughly every frame. He took the habit from web design, where he made the agent screenshot every element to check things were in bounds.
  • Roast Room = subagents with researched “brains”. Each council member (Alex Hormozi, Gary Vee, Naval, Paul Graham, Bezos, “Warren”, Codie Sanchez, Leila Hormozi, Sara Blakely, Dan Martell and others) is an agent .md file that owns one question. Hormozi’s: “will people pay, how much, and how do we close?” Each has a brain folder: hot-seat question, worldview in three sentences, eight decision rules, books, frameworks, positions, stories and quotes. Hormozi’s was built from about 123 sources and over 1.5 million of his words.
  • Subagents beat agent teams here on cost. The Roast Room originally ran as an agent team; Herk switched back to subagents because the team “was costing me more for minimal gain”. All members run the same default model.
  • The council needs options to vote on. You give it a situation plus options (or ask it to draft them); it picks the members for the question, runs opening positions, sequential one-on-one debates (sequential so others can read them), then votes and amendments, and writes a verdict plus a one-page PDF with scoring, weights, risks and an action plan. A new roster skill researches a new person and generates their cartoon avatar.
  • The Instagram account. About a month old, roughly 2K followers, 4-5 hours of his time in total. Every 4 hours it finds posts and publishes carousels; he can also send an X link from his phone (Claude Code on remote control with the Mac kept awake, or a VPS copy). Each post takes about 5-6 minutes: it reads the thread and replies, reads linked articles, downloads and clips video, generates images, layers cut-outs so faces are not covered, and renders a preview video of the carousel.
  • Verification before posting. The skill fact-checks claims from X, then re-checks when he says “post it”, because source posts are sometimes deleted or edited between drafting and posting.
  • Distribution and DM automation via Blotato. Posting runs through Blotato, which manages his X, LinkedIn, Instagram and TikTok accounts and now has ManyChat-style comment-to-DM automation. A resource-guide giveaway post drew 559 comments; commenters must follow before the DM sends. Every post is logged to a Google Sheet so topics are not repeated.

Try It

  1. Pick 20-30 of your own best-performing pieces (reels, carousels, emails) and have Claude Code extract the patterns before you write any style skill. Save the pattern bank as a reference file the skill loads.
  2. For talking-head video, get word-level timestamps (ElevenLabs or local Whisper) and add a “cut only at silence in the waveform” rule.
  3. Build a two-member council first. Write one subagent file per advisor with a single owned question, a three-sentence worldview and eight decision rules drawn from their real material, then feed it a decision with explicit options. Compare with Council, which uses different models instead of personas.
  4. For any social pipeline that sources from X, add a fact-check step and a second check at publish time, and log every published topic to a sheet the skill reads before choosing the next one.

Open Questions

  • The GitHub repo holding these skills is mentioned but not named in the source.
  • The council member captioned only as “Warren” is probably Warren Buffett; the source does not give a surname.
  • The AI video tools are captioned “key.ai or hix.ai”; the exact services are not certain. The Instagram account name is captioned “Moose Fast” and is not confirmed.
  • “Sabrina Romanov” (Blotato’s creator) is most likely Sabrina Ramonov; not confirmed from this source.
  • Whether the automated weekly analytics pull (to update the skill from post performance) is live: Herk says it is prepared but not fully set up.
  • Wolfe mentions a “reverse engineer anything” repo going viral; see REA for what the wiki has on it.